Image Spam Classification Based On Deep Learning

With the development of technology, information on the Internet becomes growing at an exploding fast speed with various contents. More specifically, pornographic images, advertising images, as well as reactionary images have made most negative effect. How to filter spam email and prevent users from bad information has become a serious problem in emergency. On the one hand, traditional methods that mainly based on shallow image information could hardly handling with images which are rising both in quantity and pattern and have been heavily limited and challenged in performance. On the other hand, considering the fact that spam emails themselves are sensitive and different from others, traditional deep learning models still have very satisfying performance on these spam images. In other words, novel models with specific structures designed need to be further researched.In this paper, we committed to solve a real problem, classifying images included in emails. To be more specific, judging the type of email according to attached image. Unlike other public spam image datasets available on the Internet, we built up our own image dataset, which was collected from spam emails. This novel dataset totally contained fifty thousand spam images roundly, separated to seven different categories,which satisfied the needs in deep learning experiments.In this paper, we proposed a new model, which was cascaded by a convolutional neural network (CNN). We also introduced a re-classification model for our final solution on this image spam emails classification task. With the introduction of transfer learning together with support vector machine (SVM), the proposed convolutional neural network shows great difference both in model structure and in parameters training process. Compared with classic convolutional neural network and normal connection convolutional neural network and support vector machine model, our model had showed better results over our spam image dataset. As to the second classifier, we first judged on the performance of the first classifier and pointed out existing problem. Then we found a local feature with strong robust and discriminate ability. We had done a set of experiments proving the affectivity in feature selection and model design.At last, we evaluated our model and gave related directions for further research and improvement.